How to train a chatbot on your help centre (2026 guide)

Most teams start here: they have a public help centre with hundreds of articles, they know most customer questions are already answered somewhere, and they want a chatbot to close the gap. Here is how to do it end-to-end without discovering the pitfalls the expensive way.

Before you crawl

Two hours of curation upfront prevents days of chatbot debugging later:

1. Prune content that no longer applies. An article about the 2023 pricing model will now be quoted by the chatbot to today’s customers.
2. Add a “Last reviewed” date to important pages. The bot doesn’t need this, but it forces your team to admit which articles have gone stale.
3. Fix titles. Google penalises weak H1s; retrievers ignore them. “FAQ” is not a title.

Configure the crawler

In Anserra, open your chatbot’s Knowledge tab, add a Website source, and configure it like this:

• URL — the help centre root, e.g. https://help.yourcompany.com
• Max depth — 3 (deep enough for category → article → sub-article)
• Max pages — start at 50, raise it only if the first crawl leaves out pages you know should be there
• Include patterns — restrict to /articles/* if you have blog posts on the same subdomain you don’t want indexed
• Exclude patterns — anything under /tags/*, pagination, and archive listings

If your help centre has a sitemap, use the Sitemap source instead. It’s more reliable than a crawl.

What actually happens after “Add source”

The crawler visits pages one at a time, respecting robots.txt and a short delay between requests. Every page is:

1. Fetched.
2. Extracted — main content only. Nav and footer are stripped.
3. Cleaned — whitespace normalised, Unicode fixed.
4. Chunked into passages of about 300–500 tokens with overlap.
5. Embedded.
6. Stored.

For most help centres, this finishes in about a minute. A source that’s still processing shows a progress bar; the chatbot won’t try to answer from half-indexed content.

Test with real questions, not made-up ones

The mistake I see teams make: they open the preview and ask “What do you offer?” That question is not what customers ask. Real customers ask specifics.

Take the twenty most common tickets from the last month and drop them into the chatbot’s preview. For each, check:

• Did it answer?
• Did it cite the right article?
• If it declined, was that because the article genuinely doesn’t exist?

The third question is the important one. A chatbot that refuses when there isn’t a good answer is doing its job.

When it declines wrongly

Two knobs to adjust:

• Minimum relevance score — if the bot rejects answers that are actually there, this is set too high.
• Chunk overlap — if answers exist in an article but the bot only finds part of the context, increase overlap so paragraphs that reference earlier sentences stay retrievable together.

Both are on your chatbot’s Retrieval tab.

Handling updates

Turn on Auto-sync on the crawl source and pick a schedule:

• Daily — for a docs site your team updates constantly.
• Weekly — the default for most teams.
• Monthly — for stable content that rarely changes.

Auto-sync re-crawls pages and compares a content hash per document, so an unchanged page costs nothing to re-embed. New articles appear in the knowledge base the same day.

What good deflection looks like

Aim for these numbers three months in:

• Resolution rate > 60% — the share of conversations where nobody had to escalate
• Thumbs-down < 8% — feedback is a leading indicator of quality drift
• Cited answers ≥ 95% — the rest are fallbacks, and that’s fine

Anserra’s Analytics tab exposes all three. Watch the “content gaps” list weekly. Every gap is a question your customers ask that your help centre doesn’t answer. Write those articles first.

Escalate on doubt

Turn on human handoff so the chatbot can pass a conversation to Slack, email, Zendesk or Intercom when either (a) it can’t find a good answer, or (b) the visitor asks for a person. The single biggest driver of customer trust is escalation that works.

The one thing that separates winners from losers

Teams that succeed with a support chatbot don’t treat it as “set and forget.” They treat it as an ongoing conversation with their own documentation. Every content gap the bot exposes gets a documentation edit, and the loop tightens over time.

If you’re ready to try it: create your first chatbot and point it at your help centre. The free plan holds up to real traffic — you only upgrade when volume genuinely grows.